A custom AI agent costs from about $8,000 to build in 2026, and most production agents land between $15,000 and $60,000, plus a monthly running cost for model usage, hosting and monitoring. The spread is wide because “AI agent” covers everything from a single-task assistant that drafts replies to a multi-step system that reads documents, calls five internal APIs and asks a human before it spends money. This guide breaks down where the money goes, what each type of AI project typically costs, what you pay every month after launch, and when buying an off-the-shelf tool is the better call.
The figures below are EqualPixels’ typical starting points for client projects, not a quote. Every project is scoped and priced before work begins, and your number depends on the drivers explained further down.
Typical AI Project Costs at a Glance
Different AI builds have very different shapes, so it helps to price them separately. These are the starting points we use when a client asks for a ballpark:
| What you are building | Typical starting point | What the starting point usually includes |
|---|---|---|
| AI workflow automation | From $2,500 | One business workflow in n8n, Make or code, with an AI step such as classification or extraction |
| MCP server | From $4,000 | A secure MCP server exposing a handful of tools from one of your systems to Claude, ChatGPT or your own agent |
| AI chatbot | From $5,000 | A chatbot trained on your content, live on your website, with handover to a human |
| AI voice agent | From $6,000 | An inbound phone agent that answers questions and books appointments into one calendar |
| Custom AI agent | From $8,000 | A single-purpose agent with two or three tools, guardrails, an evaluation set and monitoring |
| RAG knowledge assistant | From $8,000 | Retrieval over one well-organised document collection, with cited answers and access control |
The starting point is the honest minimum for something you can put in front of real users. It is not a demo. A demo that works on five hand-picked questions can be built in a few days; an agent that behaves on the five-hundredth unexpected input is where the real engineering goes.
What Actually Drives the Cost of an AI Agent?
Model choice gets most of the attention, but it is rarely what moves the build budget. These six factors do.
1. The Number of Tools and Integrations
Every tool an agent can use (search the CRM, create an invoice, check stock, send a WhatsApp message) is an integration that needs authentication, input validation, error handling and tests. An agent with two read-only tools is a small project. An agent that writes to six systems is a large one, because each write action needs permissions, limits and a way to undo mistakes.
2. How Autonomous It Is Allowed to Be
An agent that suggests an action for a person to approve is far cheaper to make safe than one that acts on its own. Full autonomy needs stricter guardrails, spending and rate limits, audit logs and more evaluation. Most businesses should start with human approval on anything that changes data or costs money, then remove the approval step for actions that prove reliable.
3. The State of Your Data
If the agent needs to answer from your documents, the quality of those documents sets the timeline. Duplicated policies, outdated PDFs and scanned images without text all have to be cleaned, converted or excluded before retrieval works. A tidy knowledge base can be indexed in days. A shared drive that has grown for ten years can take weeks.
4. Evaluation and Guardrails
A production agent needs a test set of real questions and tasks with known good outcomes, so you can measure accuracy before launch and catch regressions after every prompt or model change. It also needs guardrails: input filtering, output checks, refusal rules and limits on what each tool may do. This work typically takes 20 to 30 percent of the build and is the part cheap quotes leave out.
5. Channels and Languages
A web chat widget is the simplest channel. WhatsApp adds Meta’s template approval and messaging rules. Voice adds telephony, latency tuning and speech recognition. Each extra language, especially Arabic or Urdu alongside English, needs its own test set because accuracy rarely transfers one to one.
6. Security and Compliance
Agents that touch personal data, payments or health information need data processing agreements with model providers, region-specific hosting, redaction of sensitive fields before they reach the model, and access controls per user. None of this is optional in regulated industries, and all of it adds time.
Cost by Complexity: Three Realistic Tiers
For a custom agent specifically, most projects fall into one of three tiers.
| Tier | Typical example | Typical budget | Typical timeline |
|---|---|---|---|
| Focused agent | Drafts replies to support emails from your help centre for a person to approve | From $8,000 | 3 to 5 weeks |
| Production agent | Qualifies inbound leads, looks up the CRM, books meetings and logs every action | $15,000 to $35,000 | 6 to 10 weeks |
| Multi-agent system | Several specialised agents coordinating across documents, ERP and messaging, with human checkpoints | $35,000 to $60,000+ | 10 to 16 weeks |
These budgets assume a senior team, a fixed scope agreed upfront and weekly demos, which is how we run every engagement. If a quote is far below these numbers, ask what it leaves out. The usual answers are evaluation, monitoring and error handling.
What You Pay Every Month After Launch
The build is a one-off cost. Running the agent is not. Budget for four ongoing items.
Model Usage (Tokens)
Model providers charge per token, the small chunks of text the model reads and writes. Prices differ by model by more than an order of magnitude, and they keep falling, so check the current rates on the OpenAI pricing page and Anthropic pricing page rather than trusting any figure in an article.
A rough way to estimate: multiply conversations per month by the average tokens per conversation (system prompt, retrieved context, history and the answer together, often 4,000 to 10,000 tokens for an agent with tools). An agent that handles 3,000 conversations a month at around 6,000 tokens each uses about 18 million tokens. On a mid-tier model that is usually tens to low hundreds of dollars a month. Routing simple requests to a smaller, cheaper model and caching repeated context can cut it further.
Hosting and Vector Storage
The agent’s API, queue workers and database run on your cloud. For most small and mid-size deployments that is $50 to $300 a month. If the agent uses retrieval, the vector store adds cost too: pgvector inside an existing PostgreSQL database costs almost nothing extra, while a managed vector database such as Pinecone is billed by storage and queries.
Monitoring and Evaluation
You need tracing of every agent run (which tools it called, what it retrieved, what it answered) and a scheduled re-run of your evaluation set. Tools such as LangSmith or Langfuse, or a self-hosted equivalent, handle this. Expect a modest monthly fee or a few hours of engineering to run it yourself.
Maintenance and Improvement
Models are updated, APIs change and your business adds products and policies. Plan for a small retainer or a block of hours each month to review failed conversations, extend the test set and adjust prompts and tools. Agents that nobody looks after get worse over time, not better.
Build vs Buy: When an Off-the-Shelf Tool Is Cheaper
A custom agent is not always the right answer. Buying is usually better when:
- The job is generic, such as meeting notes, website chat on a standard help centre or email summarisation, and a SaaS tool already does it well.
- You have fewer than a few hundred conversations or tasks a month, so a per-seat subscription costs less than a build would ever save.
- You do not need the agent to act inside your own systems.
Building is usually better when:
- The agent needs to read from or write to your own CRM, ERP or database.
- The process is specific to how your business works and is a source of advantage.
- Volume is high enough that per-seat or per-conversation SaaS pricing becomes expensive.
- You need control over where data is stored and which model processes it.
A common middle path is to start with automation in a tool like n8n, prove the process works, then replace the busiest steps with a custom agent. We compare those automation platforms in n8n vs Make vs Zapier.
How to Keep Your AI Agent Budget Under Control
- Start with one workflow. Pick the task with the clearest volume and the clearest definition of “done right”. Prove it, then expand.
- Keep a human in the loop at first. Approval steps make a smaller, cheaper build safe enough to launch.
- Write the evaluation set before the prompts. Fifty real examples with known good answers will save more money than any model choice.
- Use the smallest model that passes the tests. Upgrade only the steps that fail.
- Choose boring infrastructure. Your existing database and cloud are usually enough. Agent frameworks are compared in LangGraph vs CrewAI vs OpenAI Agents SDK.
If you want a number for your own project, tell us what the agent should do. We reply within 24 hours with a scope and a fixed estimate.
FAQ
How much does a simple AI agent cost?
A focused, production-ready agent with two or three tools, guardrails and monitoring typically starts at about $8,000 with EqualPixels and takes three to five weeks. A prototype for internal testing can cost less, but it will not include the evaluation and error handling needed for real users.
What are the monthly running costs of an AI agent?
Expect model usage, hosting, monitoring and a little maintenance. For a small to mid-size deployment, model tokens are often tens to a few hundred dollars a month, hosting $50 to $300, plus monitoring and support time. High-volume voice agents cost more because speech and telephony are billed per minute.
Is it cheaper to use ChatGPT than to build a custom agent?
For individual productivity, yes. A ChatGPT or Claude subscription is far cheaper than any build. A custom agent pays off when it must work inside your own systems, follow your business rules, serve customers directly or handle volume that would need many seats.
How long does it take to build an AI agent?
Three to five weeks for a focused agent, six to ten weeks for a production agent with several integrations, and ten weeks or more for a multi-agent system. Data clean-up and third-party approvals, such as WhatsApp templates, are the most common sources of delay.
Why do AI agent quotes vary so much?
Because the hard parts are easy to leave out. A low quote often covers the prompt and a user interface but not evaluation, guardrails, monitoring, permissions or error handling. Ask every vendor how they measure accuracy before launch and what happens when the agent gets something wrong.